chore: release 0.6.3

GPT autonomous agents (OpenCode/gpt-5.5) no longer lose the live request to
history imaging — the most-recent user turn is pinned as legible text between
before/after history images, with the guard echoing it. Fixes snap/confabulation
/off-task drift; validated on gpt-5.5 (stays on task, no off-task edits).

Patch, not minor: GPT is opt-in/WIP and hidden in the dashboard, and the visible
Anthropic path is byte-identical (pin commit 08a2b0a touched only the GPT path).

Also: clarifying comment on the minCollapseTokens gate (counts imageable work
only; sub-floor histories stay fully text by design — request stays legible).

473 tests pass; tsc + build clean.

CHANGELOG.md

@@ -4,6 +4,28 @@ All notable changes to pxpipe are documented here. This project adheres to
 [Semantic Versioning](https://semver.org/) (pre-1.0: minor = features /
 behavioral changes, patch = fixes).
 
+## 0.6.3 — 2026-06-22
+
+### Fixed
+- **GPT autonomous agents no longer lose the live request to history imaging.**
+  Autonomous GPT agents (OpenCode/gpt-5.5) send one human request then a long run of
+  tool turns. The lone request is the oldest turn, so history collapse imaged it first
+  and the model lost it — confabulating the request and drifting off-task (observed:
+  editing a file instead of answering a compare question). Fix: the most-recent user
+  turn *overall* is kept as legible text, spliced between before-pin and after-pin
+  history images inside the synthetic user message; older user turns stay imaged (they
+  must not look live). The `developer` guard now echoes the request verbatim. History
+  stays imaged on both sides of the pin, so compression barely changes. Both Chat
+  Completions and Responses paths. GPT support remains opt-in/WIP (hidden in the
+  dashboard), so this does not affect the Anthropic path. (`openai-history.ts`,
+  `openai.ts`)
+
+  Cache safety (adversarially reviewed): the pin fires only when the latest user turn
+  is *inside* the collapse range — otherwise it is already native text in the kept
+  tail — so the pin's position is fixed across a run and the before/after section grid
+  stays byte-stable. An undersized before-pin remainder merges into the previous
+  section rather than emitting a sub-threshold (net-negative) image.
+
 ## 0.6.2 — 2026-06-22
 
 ### Fixed

package.json

@@ -1,6 +1,6 @@
 {
   "name": "pxpipe-proxy",
-  "version": "0.6.2",
+  "version": "0.6.3",
   "description": "Token-saving proxy for Claude Code: renders bulky context (system prompt, tool docs, old history) as dense PNGs to cut input tokens. Runs on Node and Cloudflare Workers.",
   "type": "module",
   "bin": {

src/core/openai-history.ts

@@ -281,6 +281,11 @@ export async function planGptCollapse(
   const text = joinTurns(turns, pp, rawEnd, pinIdx);
   // Floor gate in o200k TOKENS, not chars: imaging bills vision tokens and the
   // text baseline is o200k tokens, so the break-even is a token comparison.
+  // NOTE: this counts the IMAGEABLE work only (pin excluded), so a small history
+  // whose non-pin content is below the floor is left fully as text. That is correct,
+  // not a regression: the pinned request stays legible either way, and imaging a
+  // sub-floor amount of work would cost more vision tokens than it saves. Only long
+  // sessions (where the bug lived) clear the floor and collapse.
   if (!text || gptCountTokens(text) < o.minCollapseTokens) {
     return { ...base, reason: 'below_min_tokens', collapsedChars: text?.length ?? 0 };
   }